Research On Internet Traffic, Delay Properties And Prediction Model

The Internet is one of the basic infrastructures of the modern society and plays an important role in daily life. Recently, many new Internet applications, such as VoIP, IPTV and so on, have emerged. These applications usually require more strict QoS (Quality of Service). However, traditionally the Internet just provides best effort service. Therefore, how to assure QoS over the Internet becomes a challenging problem. To this end, first we need to know the state of the network. This work is achieved by network measurement technology. Then we can adjust the network according to the measured metrics and make the network running under the state expected. At the same time, the measured metric data contains laws of network running. If these laws could be uncovered through data analysis and digging technologies, then they could provide useful information for network designing, performance evaluation, QoS guaranteeing, etc. Network traffic and end-to-end delay are two concernful performance metrics. This dissertation mainly analyzes the scale property of network traffic and end-to-end delay and studies their predict models. The main contributions include:First, analyze the scale behavior of campus network traffic. The discovery of self-similarity and multifractality has brought further understanding about network running. Nevertheless, whether the multifractality exists is not consistent in past studies. There are two reasons:(1) the traffic data used in analysis is variable; (2) the analysis method used, or the rigorousness of applying the same one, is different. In this paper, we use the more believable scale analysis method, the multifractal detrended fluctuation analysis (MFDFA), to analyze campus network traffic (fine scale, below about 1 second). Our analysis provides numerical evidences for the existence of multifractality in network traffic. In addition, we investigate the source of multifractal for network traffic. We find the general two types, (1) the multifractality due to a broad probability density function (pdf) for values of the time series; and (2) the multifractality due to different long range (time-)correlations of the small and large fluctuations, can not completely explain the multifractality of network traffic.Second, study the performance of using the metabolic grey model, MGM(1,1), to predict coarse time network traffic. We find the MGM(1,1) prediction accuracy depends on the relationship between its modeling length and the length of time series periodicity. If the modeling length is far less than the periodicity length, the accuracy is satisfactory, while it is not, the accuracy deteriorates badly. We propose a combination predict model for coarse network traffic as well. The model is based on the wavelet transform, the grey theory and the chaos theory, and is denoted by WGC. In WGC, the traffic series is decomposed by an improved redundant wavelet transform into two parts, the normal part and the burst part. The normal part retains the trend and periodicity of the original series and is predicted by MGM(1,1), while the burst part is chaotic and is predicted by chaos model. Then the final predict value is reconsructured by the two predicted results. Simulation shows the WGC improves predict accuracy.Third, investigate the scale behavior of end-to-end delay. We collect delay traces by ping over both home and international paths. The DFA analysis suggests that the 2-order scale behavior of end-to-end delay is path, (collection) time and analysis scale dependent:there may be one, two or even more scale regimes, and the time correlation of each scale regime could be different (long range dependent or short range dependent). Moreover, we observe multifraclity in delay series at both fine and coarse scales as well. The multifractality source study also indicates the complexity of delay series, i.e. the two types mentioned in traffic analysis could not resolve the multifraclity in delay series, either.Finally, research the chaotic property of delay jitter series collected with small testing interval. We find the largest Lypunov exponents of these jitter series are relative small; therefor, they could be predicted well. We propose an adaptive delay predict model, neural network aided by chaos information. In this model, the chaos parameter, the embedding dimension, is used to determine the structure of neural network, and then the neural network is employed in delay (jitter) predict. Simulation results indicate the effectiveness of this model in end-to-end delay predict.